Smart meter monitoring and data mining techniques for predicting refrigeration system performance
Expert Systems with Applications: An International Journal
| Hi-index | 12.05 |
Bottom-up urban water demand forecasting based on empirical data for individual water end uses or micro-components (e.g., toilet, shower, etc.) for different households of varying characteristics is undoubtedly superior to top-down estimates originating from bulk water metres that are currently performed. Residential water end-use studies partially enabled by modern smart metering technologies such as those used in the South East Queensland Residential End Use Study (SEQREUS) provide the opportunity to align disaggregated water end-use demand for households with an extensive database covering household demographic, socio-economic and water appliance stock efficiency information. Artificial neural networks (ANNs) provide the ideal technique for aligning these databases to extract the key determinants for each water end-use category, with the view to building a residential water end-use demand forecasting model. Three conventional ANNs were used: two feed-forward back propagation networks and one radial basis function network. A sigmoid activation hidden layer and linear activation output layer produced the most accurate forecasting models. The end-use forecasting models had R^2 values of 0.33, 0.37, 0.60, 0.57, 0.57, 0.21 and 0.41 for toilet, tap, shower, clothes washer, dishwasher, bath and total internal demand, respectively. All of the forecasting models except the bath demand were able to reproduce the means and medians of the frequency distributions of the training and validation sets. This study concludes with an application of the developed forecasting model for predicting the water savings derived from a citywide implementation of a residential water appliance retrofit program (i.e., retrofitting with efficient toilets, clothes washers and shower heads).